Last year, a research nonprofit called METR ran a small experiment.

They got 16 open-source developers. People who'd been working on their own code for years. Real codebases. Real tasks.

Half the time, the developers had access to AI tools. Half the time they didn't.

Before they started, METR asked them one question. How much faster do you think AI will make you?

The developers said about 24% faster.

Then the experiment ended, and METR ran the actual numbers.

AI hadn't made them faster.

It had made them 19% slower.

Let that sink in!

A 40 point gap between what these developers believed was happening and what the clock was actually showing. They had years of experience. They knew their code inside out. And they were genuinely convinced AI was speeding them up.

The stopwatch said something else.